Leveraging Pre-Trained Diffusion Models for Unsupervised Medical Image Registration

Deformable medical image registration is critical for aligning anatomical structures in inter-patient scenarios with large variations. Traditional methods can be computationally demanding, and learning-based methods often require large datasets, limiting scalability. We propose a fully unsupervised, training-free framework that leverages pre-trained diffusion models for feature extraction. The approach removes the need for medical task-specific training or ground-truth deformation fields and supports fast inference while maintaining strong alignment accuracy. We evaluate on three datasets: FLARE22, MyOPS-bSSFP, and MyOPS-T2. The method consistently outperforms a DINOv2 pre-trained baseline and most traditional methods, and is competitive with trained models such as DiffuseMorph and UniGradICON. These results indicate that pre-trained diffusion features offer a practical alternative for medical image registration in data-limited settings.